FCL-Net: Towards accurate edge detection via Fine-scale Corrective Learning

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Multi-scale predictions are standard in edge detection, but existing methods struggle with low learning capacity in fine-level branches.
• This limitation hinders overall performance by neglecting detailed feature learning.

Purpose of the Study:

• To propose a novel network, the Fine-scale Corrective Learning Net (FCL-Net), to address the limitations in fine-level feature learning for edge detection.
• To improve the fusion performance of multi-scale predictions by enhancing fine-scale feature representation.

Main Methods:

• FCL-Net integrates semantic information from deep layers to guide fine-scale feature learning.
• Key modules include Top-down Attentional Guiding (TAG) using LSTM and Pixel-level Weighting (PW) for independent spatial location contribution.

Main Results:

• The proposed FCL-Net significantly outperforms baseline methods on benchmark datasets.
• Achieved a competitive Optimal Dataset Scale (ODS) F-measure of 0.826 on the BSDS500 dataset.

Conclusions:

• FCL-Net effectively enhances fine-scale feature learning through semantic guidance and pixel-level weighting.
• The approach demonstrates superior performance in edge detection tasks, particularly in capturing fine details.